Re: [PATCH] parisc: Remove lock code to serialize TLB operations in pacache.S

[John David Anglin <dave.anglin@bell.net>]

On 2019-04-17 5:09 p.m., John David Anglin wrote:
> On 2019-04-17 5:01 p.m., Helge Deller wrote:
>> On 17.04.19 22:58, Helge Deller wrote:
>>> On 17.04.19 22:55, Sven Schnelle wrote:
>>>> On Fri, Apr 12, 2019 at 07:12:04PM -0400, John David Anglin wrote:
>>>>> TLB operations only need to be serialized on machines with the Merced (Stretch) bus.
>>>>> The only machines in this category are L and N class, and they require a 64-bit PA 2.0
>>>>> kernel.  On these machines, we use local TLB purges in the tmpalias routines.  We don't
>>>>> need to serialize TLB purges on all other machines.  Thus, the lock/unlock code can be
>>>>> removed when CONFIG_PA20 is not defined.  Further, when CONFIG_PA20 is not defined,
>>>>> alternative patching converts the TLB purges to local purges when PA 2.0 hardware has
>>>>> been detected.
>>>>>
>>>>> Signed-off-by: John David Anglin <dave.anglin@bell.net>
>>>> I had this patch running on my C8000 and J5000 for a few days, and haven't
>>>> encountered any issues.
>>> Nah...
>>> I do have one of the critical machines (rp5470), and I tried
>>> to add the necessary code to detect and handle the TLB serialization.
>>> Until now the patch sadly doesn't work yet (hangs at boot), it still needs more work.
>>> You can find it in my git tree:
>>> https://git.kernel.org/pub/scm/linux/kernel/git/deller/parisc-linux.git/commit/?h=tlb-flush-merced-bus&id=fc6044f83cb9828ec6ebecb23a46549d9ebc518f
>> Please ignore my comments.
>> Dave's patch is OK.
>>
>> I mixed it up with the Mikulas Patocka's patch: "use per-pagetable spinlock"
>> https://patchwork.kernel.org/patch/10888249/
> I think this hunk needs fixing to handle merced case:
>
> diff --git a/arch/parisc/kernel/entry.S b/arch/parisc/kernel/entry.S
> index 5796524a3137..3e11fdd9d3af 100644
> --- a/arch/parisc/kernel/entry.S
> +++ b/arch/parisc/kernel/entry.S
> @@ -48,14 +48,9 @@
>      .level 2.0
>  #endif
>  
> -    .import        pa_tlb_lock,data
> -    .macro  load_pa_tlb_lock reg
> -#if __PA_LDCW_ALIGNMENT > 4
> -    load32    PA(pa_tlb_lock) + __PA_LDCW_ALIGNMENT-1, \reg
> -    depi    0,31,__PA_LDCW_ALIGN_ORDER, \reg
> -#else
> -    load32    PA(pa_tlb_lock), \reg
> -#endif
> +    .macro  load_process_pte_lock reg
> +    mfctl        %cr25,\reg
> +    addil        L%((PAGE_SIZE << PGD_ALLOC_ORDER) - PAGE_SIZE),\reg
>      .endm
>  
>      /* space_to_prot macro creates a prot id from a space id */

Another option might be to remove pa_serialize_tlb_flushes here:
+    if (unlikely(pgd == swapper_pg_dir) || pa_serialize_tlb_flushes)
+        return &pa_tlb_flush_lock;
This might be the better solution.  You have serialized the TLB purges using the pa_tlb_flush_lock.

Dave

-- 
John David Anglin  dave.anglin@bell.net